Truth Recognition
How do you know what you know is true?
Andrej Karpathy spent four hours building a convincing argument with an LLM. Then he asked it to argue the opposite. It demolished everything. Four hours of conviction, undone in seconds.
The failure was not the model. The failure was the control system.
| Without Truth Recognition | With Truth Recognition |
|---|---|
| Convinced by last argument heard | Anchored before the argument starts |
| Pattern-match to plausible | Test against verifiable |
| Accept at face value | Demand the mechanism |
| Beliefs blow in the wind | Beliefs update through evidence |
Control System
Truth recognition is a VVFL — the same feedback loop that runs everywhere. Four components. Remove one and the loop breaks.
| VVFL Role | Component | Instrument | What It Does | Without It |
|---|---|---|---|---|
| Setpoint | Identity | Who you are | Reference for all evaluation | Gauge reads anything — Karpathy's lesson |
| Gauge | Pattern Recognition | What you observe | Detect signals in noise | Noise looks like signal |
| Gauge | Questioning | What you probe | Test what you observe | Observations go unexamined |
| Controller | Critical Thinking | How you evaluate | Compare against evidence | Accept first plausible answer |
| Controller | Truth-Seeking Protocol | Mechanical gate | Forces the question when you forget | Drift under cognitive load |
| Feedback | Verified Truth | What you now know | Updates identity for next cycle | Loop reprograms rather than refines |
IDENTITY (setpoint)
↓
OBSERVE (pattern recognition) → PROBE (questioning)
↓
EVALUATE (critical thinking) → ENFORCE (protocol gate)
↓
TRUTH (verified) → updates IDENTITY
↓
The loop compounds
The setpoint is the key. Wayne Smith's question-based learning is truth recognition in action — the coach asks, the player discovers, ownership comes from solving the problem yourself. Without the identity anchor (who are we, what do we stand for), the discovery has no frame.
Three Scales
The same control system operates at three scales. The meta is the pattern. The matter is which layer you are operating at.
| Scale | Identity Layer | Truth Infrastructure | Verification | Example |
|---|---|---|---|---|
| Personal | Character, values, pepeha | Self-awareness, reflection, journaling | Honest mirror — what do you do when no one watches? | Karpathy at the keyboard |
| Social | Culture, community, belonging | Shared standards, protocols | Peer challenge — the crew holds each other accountable | Rennie's Chiefs knowing who they play for |
| Cryptographic | DIDs, ZKPs, proof of personhood | Blockchain, DePIN | Mathematical proof — verification without trust | Identity that travels across networks |
Each scale solves a different problem. Personal truth-seeking fails at scale (one person's conviction does not constitute shared reality). Social truth-seeking fails without personal integrity (groupthink replaces truth). Cryptographic truth-seeking fails without human meaning (proof of personhood proves nothing about character).
The full stack: personal integrity grounds social standards which are verified by cryptographic infrastructure.
Evidence Ladder
Not all truth claims are equal. From the truth-seeking protocol:
| Level | Type | Weight | Action |
|---|---|---|---|
| 1 | Observed failure — logs, errors, complaints | High | Act |
| 2 | Measured inefficiency — metrics, benchmarks | Medium | Investigate |
| 3 | Architectural observation — "two systems exist" | Low | Document |
| 4 | Theoretical concern — "this might cause..." | None | Note and move on |
Level 3-4 evidence is insufficient to justify change. This is where most false conviction lives — the pattern that looks true because it is plausible, not because it is verified.
Practice Protocol
- Anchor first — Before seeking truth on any topic, state what you believe and why. Write it down. This is your setpoint. Without it, you are Karpathy at hour four.
- Observe before concluding — Run pattern recognition. What are you actually seeing? Cross-domain: where have you seen this shape before?
- Probe before accepting — Run questioning. What would have to be true for this to be false? What is the strongest case against?
- Evaluate against evidence — Run critical thinking. Source check. Evidence check. Logic check. Bias check. What level on the evidence ladder?
- Update, don't replace — Verified truth refines identity. It does not demolish it. If one argument flips your entire worldview, the setpoint was never yours.
The Shadow
Paranoid skepticism. Refusing to update because "nothing can be trusted." Treating all sources as equally suspect — the conspiracy trap. The opposite failure: naive acceptance of whatever sounds authoritative. Both are control system failures. Skepticism without a setpoint is paralysis. Acceptance without a controller is drift.
The diagnostic: "What would change my mind?" If the answer is "nothing" — that is faith, not truth recognition. If the answer is "anything convincing enough" — there is no setpoint.
By Archetype
| Archetype | Truth Style |
|---|---|
| Philosopher | Seeks structural truth — what must be true regardless of perspective? |
| Realist | Seeks evidential truth — what can be measured and verified? |
| Engineer | Seeks operational truth — does this work under load? |
| Coach | Seeks emergent truth — what does the player discover when asked? |
Context
- Identity — The setpoint. Without it, the loop has no reference.
- Pattern Recognition — The gauge. What you observe.
- Questioning — The probe. What you test.
- Critical Thinking — The controller. How you evaluate.
- Truth-Seeking Protocol — Mechanical enforcement. Forces the question.
- The Truth — Truth as infrastructure. Verification at scale.
- Blockchain Identity — Identity made portable, sovereign, verifiable.
- Control System — The VVFL pattern this page implements.
- Persuasion — The force truth recognition defends against and deploys.
- Behavioural Biases — The predictable failures truth recognition catches.
Links
- Karpathy — LLM Persuasion — Four hours of conviction demolished in seconds
- AI Superhuman Persuasion — Altman's prediction and the evidence
- Algorithmic Self — How AI reshapes identity, introspection, and agency
- Urgency of Interpretability — What you cannot interpret, you cannot verify
Questions
If Karpathy — one of the sharpest minds in AI — had his conviction flipped in seconds, what makes you think yours would survive?
- The control system has four components (setpoint, gauge, controller, feedback) — which one is weakest in your own truth-seeking practice?
- Personal truth-seeking fails at scale, social truth-seeking fails without integrity, cryptographic fails without meaning — which failure mode is your organisation most exposed to?
- The evidence ladder says Level 3-4 evidence is insufficient to justify change — how much of what you currently believe was adopted at Level 4?
- Wayne Smith asks instead of telling. The player discovers truth through the question. What happens when the question comes from an AI that can lead you anywhere?